Are ChatGPT logs admissible as evidence in US courts?

Yes. In U.S. V. Rinderknecht (the Palisades Fire trial, June 2026), Judge Anne Hwang admitted ChatGPT text conversation logs under existing Federal Rules of Evidence for electronic communications, while excluding AI-generated images as too prejudicial. The retrial is scheduled for October 19, 2026.

What legal rules govern AI conversation logs as evidence?

Courts have applied FRE 901 (authentication), FRE 803(6) (business records exception), FRE 807 (residual hearsay), and FRE 801(d)(2) (admission by party-opponent) to AI logs. The user's prompts are treated like any electronic message; the AI's responses raise an unresolved 'declarant problem' because no human authored them.

What does the Palisades Fire case mean for corporate AI governance?

It means AI conversation logs are now potentially discoverable evidence. Companies must audit which AI tools employees use, set retention and deletion schedules, establish legal-response protocols, and train staff that chats can be subpoenaed and used against them or the firm.

Does the EU AI Act require record-keeping of AI conversations?

Yes. Article 12 of the EU AI Act mandates automatic event logging for high-risk AI systems, with logs retained at least six months. Article 11 requires technical documentation, and Article 50 imposes chatbot transparency duties. Most obligations take full effect August 2, 2026.

Can a company be liable for what its AI chatbot says?

Yes. The Air Canada chatbot case established that companies can face liability for outputs their AI systems produce, extending traditional product-liability principles to AI-generated content and creating pressure to govern chatbot behavior.

ChatGPT Logs as Evidence: What the Palisades Fire Trial Means for AI

On June 26, 2026, a federal jury in the Central District of California deadlocked 10-2 for acquittal in the arson trial of Jonathan Rinderknecht, the man accused of starting the Palisades Fire that burned through Los Angeles in January 2025. The mistrial made headlines, but the quieter precedent was already set: prosecutors had admitted ChatGPT conversation logs as evidence to prove intent and premeditation, and Judge Anne Hwang let them in.

AI conversation logs are now admissible evidence in an American criminal court, treated with the same procedural deference as email. That single shift rewrites what every company must do about AI governance.

TL;DR

U.S. V. Rinderknecht is the first major U.S. Criminal case where ChatGPT logs carried real evidentiary weight in court. Prosecutors used the transcripts to show the defendant had researched fire dynamics and expressed a "yearning for revenge on society," according to The Verge.

Judge Hwang admitted the text logs but excluded AI-generated images as "very, very prejudicial." The retrial is set for October 19, 2026. For corporate AI governance, the message is blunt: treat AI conversation logs like email, because courts already do.

Key takeaways

ChatGPT text logs were admitted under existing Federal Rules of Evidence for electronic communications, not a new AI-specific doctrine.
AI-generated images were excluded as prejudicial, establishing an early text-vs-image split in AI evidence law.
The user's prompts are treated like admissions by a party-opponent; the AI's responses raise an unresolved "declarant problem."
OpenAI produced the logs through subpoena and voluntary compliance, setting a de facto disclosure pattern.
The EU AI Act's Article 12 logging requirements take full effect August 2, 2026, creating a parallel record-keeping regime.
Companies must now audit AI tool use, set retention schedules, and prepare for AI-log discovery.

What Did the Court Actually Admit?

Prosecutors charged Rinderknecht with destruction of property by fire, arson affecting interstate commerce property, and timber set afire. To prove intent, they introduced ChatGPT logs showing he had asked the model about fire dynamics, forest burning patterns, and climate change, and had expressed admiration for Luigi Mangione, the man convicted of killing UnitedHealthcare CEO Brian Thompson (CNN).

Judge Hwang's ruling was a split decision. She admitted the text conversation logs as evidence of state of mind and premeditation. She excluded the AI-generated images Rinderknecht had created, calling them "very, very prejudicial" (The Verge).

That distinction matters. Courts appear more willing to admit the written record of a human's prompts to an AI than the visual artifacts the model produced in response.

The admission did not rely on any novel AI doctrine. Prosecutors handled the logs with the same authentication and chain-of-custody standards applied to email and other electronic records.

Richard Salgado, a lecturer at Stanford Law School's Cyberlaw program and former DOJ computer crime prosecutor, told the Cybersecurity Law Report that prosecutors are giving generative AI chats "the respect given to email and other nonpublic content."

How Does AI Evidence Law Work Under the Federal Rules?

The Palisades case forced courts to apply three familiar hearsay frameworks to a new kind of record: FRE 803(6) (business records exception), FRE 807 (residual exception), and FRE 801(d)(2) (admission by party-opponent). Each one strains differently when applied to AI logs.

Framework	What it covers	How it applies to AI logs	Open problem
FRE 803(6)	Business records kept in the regular course of business	OpenAI's logs are contemporaneous and kept as a regular business practice	The "made by a person with knowledge" requirement for AI outputs
FRE 801(d)(2)	Admissions by a party-opponent	The user's prompts function like any electronic message from the party	Does not cover the model's responses
FRE 807	Residual hearsay	Catch-all for trustworthy statements not covered elsewhere	Courts rarely reach this if 803(6) or 801(d)(2) suffice

The user's prompts are the easy part. They read like any message a party sends, and they qualify as admissions by party-opponent under FRE 801(d)(2). The AI's responses are the hard part.

Hearsay requires an out-of-court statement by a human declarant. A language model is not a human, has no personal knowledge, and has no intent. Some scholars argue AI outputs are simply not hearsay because they fail the definitional requirement of a human "statement."

Others say the human programmers' design choices could make the outputs attributable to them. No appellate court has resolved this yet.

Authentication under FRE 901 required the prosecution to show the logs were what they claimed to be. The Palisades team satisfied this through OpenAI's cooperation with subpoenas plus forensic testimony establishing authenticity (AIFCF analysis). In future cases where an AI provider is less cooperative or a defendant disputes completeness, authentication will be harder.

Why This Case May Be Sui Generis

Not everyone agrees the Palisades ruling generalizes. Royal Oakes, a former prosecutor commenting on the mistrial, told the New York Post that "the first trial sounded like a case about the defendant's personality" and that "the retrial has to be a case about fire science," because "jurors don't convict because they think somebody is angry, eccentric or resentful."

Neama Rahmani, a former Assistant U.S. Attorney, speculated the deadlock reflected "jury nullification" or jurors blaming government agencies rather than doubts about the AI evidence (AOL/Fox News). The case had no direct physical evidence like video, DNA, or fingerprints.

The procedural posture also limits precedential value. A mistrial means no appellate court has reviewed Judge Hwang's evidentiary rulings. Trial court admissibility decisions carry less weight than appellate ones, especially when the trial did not end in a verdict (Legal Experts).

The retrial, or any subsequent appeal, will matter more than the first trial's rulings.

That said, the practitioner consensus treats the case as operationally significant regardless. McCarter & English's analysis notes that law enforcement and civil litigants are already treating AI conversation records as a legitimate evidence category, with the same authentication standards as other electronic communications (McCarter & English).

Rob T. Lee, Chief AI Officer at the SANS Institute, called it "one of the first US cases where ChatGPT logs carried real evidentiary weight in court," adding "it won't be the last" (ZealTyro).

What This Means for Corporate AI Liability

The Palisades case is criminal, but the same evidentiary logic applies in civil and corporate contexts. If a ChatGPT log can prove a defendant's premeditation in an arson trial, it can prove an employee's intent in a securities fraud case, a product liability dispute, or an IP theft claim.

Lee's description of the Rinderknecht logs as "someone workshopping their alibi in real time" applies just as well to an employee workshopping a cover story on a company-connected AI tool.

The liability surface has two sides. On the user side, employees generate retainable records every time they prompt a model. On the provider side, AI companies face questions about what they retain, what they disclose, and what duty they have to preserve logs for law enforcement.

OpenAI produced Rinderknecht's logs through subpoena and voluntary compliance. Companies that retain too little may be unable to defend themselves; companies that retain too much invite privacy litigation and regulatory scrutiny.

The Air Canada chatbot case already established that companies can face liability for the outputs their AI systems produce (Pinsent Masons). Add the Palisades precedent, and the picture is clear: AI product liability now spans both what your chatbot says and what your users say to it.

How Should Corporate AI Policy Respond?

Effective corporate AI governance after Palisades has four concrete components.

1. Audit what's in use. Map every AI system employees touch, including consumer tools like ChatGPT used on personal devices for work tasks. Shadow AI use is widespread and largely ungoverned (Cybersecurity Dive. You cannot govern what you have not inventoried.

2. Set retention and deletion schedules. Treat AI conversation logs with the same rigor as email. Implement defensible deletion schedules, audit trails, and clear policies on what is retained and for how long. The goal is a record-keeping posture that survives discovery without exposing the firm to over-retention claims.

3. Build a legal-response protocol. Establish procedures for validating subpoenas, protecting privileged content, and notifying affected employees when AI logs are sought. The protocol should specify who can authorize disclosure and how chain-of-custody is documented.

4. Train employees. Workers need to understand that chats with an AI tool are not ephemeral. They can be subpoenaed, authenticated, and read aloud in court. A short, specific training beat beats a long acceptable-use policy nobody reads.

Governance without enforcement is theater. The Samsung ChatGPT data leak showed what happens when employees paste sensitive source code into a consumer chatbot with no guardrails (AI Alert). The Palisades case shows the other failure mode: chats that employees assume are private becoming the centerpiece of a prosecution.

How Does the EU AI Act Change the Record-Keeping Picture?

While U.S. Evidence law develops case by case, the EU has built a statutory record-keeping regime. The EU AI Act takes full effect for most providers on August 2, 2026 (Eurocomply).

Article 12 mandates automatic event logging for high-risk AI systems, with logs retained at least six months by providers and deployers (EU AI Act Service Desk). Article 11 requires technical documentation covering training data, algorithms, and system architecture (Article 11).

Article 50 imposes transparency duties on chatbot providers: users must be informed they are interacting with an AI unless it is obvious from context.

That transparency duty has a direct evidentiary consequence. If a user is clearly told they are talking to an AI, their subsequent prompts may be treated differently than communications with a human interlocutor, which affects both authentication and hearsay analysis.

Companies operating globally now navigate two regimes at once: U.S. Courts admitting AI logs under common-law evidence rules, and EU regulators mandating logging and documentation by statute.

In the U.S., the landscape remains fragmented. The Trump administration's December 2025 executive order "Eliminating State Law Obstruction of National Artificial Intelligence Policy" signaled a federal-centralization intent (White House).

The FTC has pursued deceptive-AI enforcement and opened an inquiry into companion chatbots (Lawfare). The NIST AI Risk Management Framework offers voluntary guidance. There is still no comprehensive federal AI statute, which means the evidentiary questions will keep being resolved one case at a time.

What This Means for You

Three moves you can make this quarter.

First, run an AI tool inventory across your organization, including personal-device use for work tasks. If you cannot produce a list of which AI systems touch your data, you cannot respond to a subpoena competently.

Second, align your AI log retention policy with your email retention policy. Same schedules, same deletion discipline, same legal-hold mechanics. The Palisades ruling gives you the justification to treat the two categories identically, and your legal team will recognize the framing.

Third, add a one-paragraph AI-evidence clause to employee AI training. The message is simple: anything you type into a chatbot can be read back to you in a deposition. Rob T. Lee called AI logs a "new category of evidence." Your employees should know that before they type, not after.

The retrial of U.S. V. Rinderknecht begins October 19, 2026. Whatever the verdict, the admissibility precedent is already doing work in courtrooms and boardrooms. The companies that treat AI conversation logs as ordinary business records now will be the ones that survive discovery later.

ChatGPT Logs Are Now Evidence. The Palisades Fire Trial Rewrites AI Governance.